The number you computed is nearly 1/e. Basically, (1-1/n)^n = exp(n log (1-1/n)) = exp(n (-1/n + 1/(2n^2) + O(1/n^3) )) = exp(-1 + 1/(2n) + O(1/n^2) ) = 1/e - 1/(2en) + O(1/n^2) and so your answer should agree with 1/e to, say, the first forty digits or so.

(Incidentally, Maple choked when I plugged in (1-2^-128)^(2^128), because it tried to evaluate it as a rational number, the numerator and denominator of which would have well over 2^128 digits.)

Very interesting. I had not worked out or seen the math for this. I don't doubt that such a number could exist. I do question the speed with which one could find the number and the utility once it's found.

How to find it: do MD5 on some random input. Then do MD5 on that output, and repeat until it converges. The problem is that you could end up in a cycle, instead of at a fixed point.

The speed with which it can be found: I'm going to wave my hands and claim this is in "Analytic Combinatorics" by Flajolet and Sedgewick. Seriously, though, under the assumptions we've been throwing around here this is a "random mapping" and these are reasonably well-studied objects.

It would be interesting to try to map the group properties of MD5-space. Every 128-bit number would either be in a slide to a cycle or in a cycle (a fixed point is a cycle of size 1).

This still doesn't change the huge-normousness of the task, though.

No need to worry about cycles from using one output as the next input; just iterate over all possible 128-bit inputs, in any order. Unless you're trying to leverage some known deviation of MD5 from being a true random oracle, each input is just as likely as any other to be an identity input.